Artificial intelligence synthesized face swapping detection model using unified data sets
Today’s image generation technology can generate high-quality face images, and it is not easy to recognize the authenticity of the generated images through human eyes. Due to the rise of image generation technology based on deep learning, software related to image generation is used widely, includin...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | http://eprints.utem.edu.my/id/eprint/28296/1/Artificial%20intelligence%20synthesized%20face%20swapping%20detection%20model%20using%20unified%20data%20sets.pdf http://eprints.utem.edu.my/id/eprint/28296/2/Artificial%20intelligence%20synthesized%20face%20swapping%20detection%20model%20using%20unified%20data%20sets.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Today’s image generation technology can generate high-quality face images, and it is not easy to recognize the authenticity of the generated images through human eyes. Due to the rise of image generation technology based on deep learning, software related to image generation is used widely, including some popular face-swapping software. If misused, it will directly affect forensics and security-related industries. As an essential branch of computer security, image forensics technology also needs to be improved with the development of image forgery technology. This study aims to improve deepfake detection, a face-swapping forgery, by absorbing the advantages of deep learning technologies. In order to solve the problem of poor detection performance on cross data sets, this study generates unified data sets from multiple sources using spatial enhancement technology to obtain approximately four million images, 36 times the size of the original data set, and was proved effective with traditional feature methods. Taking the advantages of ResNet and Inception networks, DeepfakeNet architecture composed of 32 parallel branches and 20 network layers is proposed as the deepfake detection model with FLOPs of 2.05 × 109 and parameters of 10.87 × 106. To further improve the proposed DeepfakeNet model, a univariate method is used to obtain the ideal model values of hyperparameters, including batch size, epochs, dropout, learning rate, and sample ratio. Accuracy of 98.69%, loss value of 3.42% and AUC of 0.96 are achieved. The evidence of this study shows that the proposed DeepfakeNet has significantly improved over the mainstream methods in terms of loss value, accuracy, AUC, FLOPs, and parameters. |
|---|